Rocky Duan: from OpenAI to Covariant and building foundation models for robotic automation

Warehouse automation requires moving beyond rigid programming toward AI systems that learn from data and trial-and-error. Covariant’s approach utilizes deep reinforcement learning to enable robots to manipulate diverse, unstructured objects, effectively addressing the challenges of warehouse pick-and-place tasks. Foundation models, trained on internet-scale data, provide a mechanism for robots to acquire common-sense reasoning, allowing for rapid adaptation to long-tail scenarios like damaged packaging or unexpected object layouts. Beyond logistics, the frontier of AI robotics involves creating highly reliable systems capable of autonomous decision-making and potential expansion into automating complex scientific laboratory experiments. As CTO and co-founder of Covariant, Rocky Duan emphasizes that the ability to learn from past experience and adapt instantly to environmental changes is essential for scaling intelligent robotic systems in real-world commercial environments.